Using NIR/SWIR hyperspectral camera mounted on a UAV to assess grapevine water status in a variably irrigated vineyard

Abstract

Context and purpose of the study – Vineyards face climate change, increasing temperatures, and drought affecting vine water status. Water deficit affects plant physiology and can ultimately decrease yield and grape quality when it is not well managed. Monitoring vine water status and irrigation can help growers better manage their vineyards. However, when field measurements, such as stem water potentials (SWP), can be precise, they are time-consuming. In addition, they do not allow for easy assessment of spatial variability, which is a critical factor for water status management. Remote sensing tools can help map plant water status in space and time and streamline data acquisition over whole vineyards several times during the season. In this project, we monitored a variably irrigated vineyard several times during the season with a hyperspectral NIR/SWIR camera mounted on a UAV.

Material and methods – We worked in a Cabernet Sauvignon vineyard in the San Joaquin Valley of California equipped with an automated irrigation system. We created forty-eight independent watering zones and applied twelve different amounts of water replicated four times in a randomized block scheme. Water amounts were fractions of the grower allocation and applied as sustained and regulated deficit irrigation strategies. Hyperspectral images in 112 bands from 900 nm to 1700 nm were collected using a UAV every two weeks from June to harvest. Contemporarily, we measured vine water status through SWP, stomatal conductance (g_s) and net assimilation (A_N). For the analysis, the images were segmented to extract the canopy signal and converted to reflectance, then used to predict the field water status measurements using machine learning models. Models were evaluated using coefficients of determination (R²), and root mean square error (RMSE). Feature importance was also computed to determine the importance of each band in the model.

Results – Field measurements of stem water potential ranged from -2.0 to -1.14 MPa. The canopy signal was segmented from the soil background using a classifier with an accuracy of 99.7%. We tested random forest, gradient boosting machine, and support vector machine algorithms in a preliminary analysis to predict SWP values. The most performant model was the random forest, and it was able to predict SWP values with an R² of 0.6 and an RMSE of 0.1 MPa as assessed in a 5-fold cross-validation procedure. The most important bands for model prediction were 1146 nm, 1153 nm, 1321 nm, 1363 nm, and 1434 nm, all situated in water absorption domains. These promising results demonstrate that SWIR images can monitor the field’s vine water status and inform irrigation management with high resolution.